OBLITERATUS Unchains gemma-4-E4B-it-OBLITERATED v3

    
        By vramkickedin    
     | 
    
            April 27, 2026 at 4:37 pm        
    
     | 
    
        2 min read

OBLITERATUS recently released v3 of gemma-4-E4B-it-OBLITERATED, a modified version of Google’s Gemma 4 designed to eliminate hard refusals. The 4B parameter model strips away built-in content filters while preserving core reasoning functions.

Local AI researchers created this update to enable unrestricted prompt testing and system tuning. The release resolves earlier compatibility errors that previously caused missing data and degraded output quality.

Model Size: from 5.34GB & VRAM GPU: 8GB required

Guardrails removed at the layer level

Achieves zero percent hard refusal across standard evaluation sets.
Retains all 720 model tensors through corrected weight mapping.
Includes tuned generation parameters to minimize output repetition.
Supports modern desktop frameworks and recent mobile inference applications.

Developers managing private documentation workflows gain consistent generation that adapts directly to provided instructions. Agencies handling sensitive client data can run the entire pipeline offline without external network dependencies.

Architecture fixes and testing notes

The creation process used an automated agent requiring fewer than ten human commands to repair corrupted projections. Early builds triggered mathematical errors because the shared weight system duplicated calculations, but v3 applies corrections exactly once to prevent data loss.

Builders warn that smaller parameter counts naturally limit complex reasoning, so operators should adjust temperature settings to control tangential responses. OBLITERATUS noted the v3 improvements directly in the project page, stating